DATA2002 Module 4 Report

library(tidyverse)
library(knitr)
library(janitor)
library("readxl")
library(ggfortify)
library(GGally)
library(qtlcharts)
library(leaps)
library(sjPlot)
library(pheatmap)
library(partykit)
library(rpart)

1. Introduction

Excess weight, especially obesity, has become an epidemic in the 21st century and resulted in many significant health and economic consequences for the global population (Stein and Colditz, 2004). In Australia, this epidemic has also spread as 2 in 3 adults are classified as overweight or obese (Australian Institute of Health and Welfare). Researches has shown that this epidemic is more common in males than females and hence, BYU Human Performance Research Center has collected data from 250 men of various age and obtained estimates of the percentage of body fat through underwater weighing and various body circumference measurements (Rahman and Harding, 2013; DASL, n.d.). As body fat percentage is difficult to calculate in real life, the value for body fat percentage was derived from body density using the Siri’s 1956 equation.

1.1 Sampling method and potential biases

There were not many information given with regards to the sampling method. However, from looking at the dataset, there is gender biase as the epidemic question is one related to both gender, yet only male were involved in the sample. This suggests that any analysis based on this dataset cannot be applied to the whole population but only the male population.

1.2 Data import, processing and cleaning

data = read.delim("bodyfat.txt") %>% janitor::clean_names()
data = data %>%
  mutate(bmi = (data$weight/(data$height ^ 2)) * 703,
         overweight = case_when(
          bmi >= 25 ~ 1,
          bmi < 25 ~ 0))
<<<<<<< HEAD
data$overweight = as.character(data$overweight)
=======
data$overweight = as.numeric(data$overweight)
>>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419
#colnames(data)
data_bmi = data[-c(1:2,4:5,18)]
data_bf = data[-c(1,3:5,17:18)]
data_density = data[-c(2:5,17:18)]
#glimpse(data)

1.3 Data Visualisation

1.3.1 Obesity and Age

ggplot(data, aes(x=age,fill=overweight)) + geom_bar() +scale_fill_brewer(palette = "Paired") +labs(x='Age',y='Count', title = 'Severness of Obesity Across Different Age Groups',caption='Source: DASL Data') +theme(plot.title = element_text(hjust = 0.5, face = 'bold'))+theme(axis.text.x = element_text(angle = 50, hjust = 1))+scale_fill_discrete(name = "Overweight", labels = c("Not Over-Weight", "Over-Weight"))

## Scale for 'fill' is already present. Adding another scale for 'fill',
## which will replace the existing scale.

<<<<<<< HEAD

=======

>>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419

From looking at the graph, it can be observed that most of the sample comes from the age group between 40-60 and the number of people who appear to be over-weight seems to be distributed evenly across the population.

1.3.2 Body Measurements

boxplot(data$neck,data$chest,data$abdomen,data$waist,data$bicep,data$forearm,data$wrist,main = "Upper Body Measurements",names = c("Neck","Chest","Abdomen","Waist","Bicep","Forearm","Wrist"))

<<<<<<< HEAD

boxplot(data$hip,data$knee,data$thigh,data$ankle,main = "Lower Body Measurements",names = c("Hip","Knee","Thigh","Ankle"))

=======

boxplot(data$hip,data$knee,data$thigh,data$ankle,main = "Lower Body Measurements",names = c("Hip","Knee","Thigh","Ankle"))

>>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419

From the two boxplots, the body measurement dataset appears relatively symmetrical and contains a few outliers. This may propose some problems in the later stages of the analysis.

1.4 Analysis Approach

This report aims to determine an alternative method to determine “over-weight” individuals oppose to body fat percentage and the two alternative methods considered are:

BMI
Body Density

The analysis will first begin through determining how much variation in body fat percentage can be explained by simply body measurements and the number of measurements that is significant in building an accurate prediction model to examine the ease of calculation.

Then similar analysis will be conducted on BMI and Body Density where the end result will be compared together to determine which method can be explained the best using simple body measurements and offers easier and simpler interpretation.

Lastly, a binary indicator will be added to differentiate the sample into over-weight individuals (Yes) and non-over-weight individuals (No) using body-fat percentage as the guiding criteria. A logstic regression is run on the binary indicator with BMI, body density and age for a simpler model to determine the odds of an individual being obesed.

2. Analysis

Due to the increasing consumptions of fast food and the increasing convenience of food deliveries, concerns about obesity level is rising throughput the world and has reached a new high. This increasing concern has lead to an increasing need to measure obesity accurately and percentage body fat is arguably the most accurate measure by far. However, the calculation of body fat is difficult and many has switched to Body Mass Index (BMI) for simpler calculation. This section is looking at how much variation that simple body measurements can explain in the three methods of interest - body fat percentage, BMI and body density.

<<<<<<< HEAD

2.1 Body Fat Percentage

2.1.1 Data Visualisation

=======

qtlcharts::iplotCorr(data_bf)

## Set screen size to height=700 x width=1000

=======

2.1 Body Fat Percentage

2.1.1 Data Visualisation

qtlcharts::iplotCorr(data_bf)

## Set screen size to height=700 x width=1000

>>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419

Based on the interactive correlation matrix above, it can be seen the level of correlation differs quite drastically between the variables and the backward variable selection method is adopted.

2.1.2 Multiple Regression and Variable Selection

bf_lm = lm(pct_bf~.,data=data_bf)
summary(bf_lm)

## 
## Call:
## lm(formula = pct_bf ~ ., data = data_bf)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -9.8684 -2.9088 -0.1904  3.0491 11.1421 
## 
## Coefficients: (1 not defined because of singularities)
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  2.20340    6.83392   0.322  0.74742    
## neck        -0.45612    0.23034  -1.980  0.04882 *  
## chest       -0.13005    0.09197  -1.414  0.15866    
## abdomen      1.03299    0.07638  13.524  < 2e-16 ***
## waist             NA         NA      NA       NA    
## hip         -0.33000    0.12768  -2.585  0.01034 *  
## thigh        0.08793    0.13395   0.656  0.51217    
## knee        -0.13537    0.22744  -0.595  0.55227    
## ankle        0.05505    0.21751   0.253  0.80041    
## bicep        0.17762    0.17029   1.043  0.29798    
## forearm      0.19468    0.20718   0.940  0.34834    
## wrist       -1.52499    0.50529  -3.018  0.00282 ** 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 4.341 on 239 degrees of freedom
## Multiple R-squared:  0.737,  Adjusted R-squared:  0.726 
## F-statistic: 66.98 on 10 and 239 DF,  p-value: < 2.2e-16

Using the individual p-value method, the varaibles that need to be dropped are chest, waist, thigh, knee,ankle, bicep, forearm with ankle being the first to drop down due to its high p-value. However, to double check, the AIC criterion will also be considered.

bf_step_back = step(bf_lm, direction = "backward",trace = FALSE)
summary(bf_step_back)

## 
## Call:
## lm(formula = pct_bf ~ neck + chest + abdomen + hip + bicep + 
##     wrist, data = data_bf)
## 
## Residuals:
##    Min     1Q Median     3Q    Max 
## -9.668 -2.889 -0.361  3.210 11.148 
## 
## Coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  1.52703    6.63727   0.230 0.818232    
## neck        -0.39650    0.22234  -1.783 0.075783 .  
## chest       -0.12810    0.08992  -1.425 0.155562    
## abdomen      1.01805    0.07431  13.700  < 2e-16 ***
## hip         -0.28758    0.09232  -3.115 0.002060 ** 
## bicep        0.26094    0.15160   1.721 0.086469 .  
## wrist       -1.55084    0.45510  -3.408 0.000767 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 4.32 on 243 degrees of freedom
## Multiple R-squared:  0.7353, Adjusted R-squared:  0.7287 
## F-statistic: 112.5 on 6 and 243 DF,  p-value: < 2.2e-16

Based on the backward selection model, the fitted model has become:

\[ \hat{Body - Fat} = 1.52 -0.3965Neck - 0.128Chest + 1.01805Abdomen -0.28758Hip + 0.26Vicep -1.55084Wrist \]

2.1.3 Check Assumptions

Finally, to check assumption, we perform the ggfortify function.

par(mfrow=c(1,2))
plot(bf_step_back,which=1:2) + theme_bw()

## NULL

The QQ plot shows a straight line which indicates that the normality assumption is reasonable. However, the residuals vs fitted plot shows a slight variation; but given that body fat is hard to predict, this is acceptable.

2.1.4 Final fitted model

relbf <- function(fit,...){
  R <- cor(fit$model)
  nvar <- ncol(R)
  rxx <- R[2:nvar, 2:nvar]
  rxy <- R[2:nvar, 1]
  svd <- eigen(rxx)
  evec <- svd$vectors
  ev <- svd$values
  delta <- diag(sqrt(ev))
  lambda <- evec %*% delta %*% t(evec)
  lambdasq <- lambda ^ 2
  beta <- solve(lambda) %*% rxy
  rsquare <- colSums(beta ^ 2)
  rawwgt <- lambdasq %*% beta ^ 2
  import <- (rawwgt / rsquare) * 100
  import <- as.data.frame(import)
  row.names(import) <- names(fit$model[2:nvar])
  names(import) <- "Weights"
  import <- import[order(import),1, drop=FALSE]
  dotchart(import$Weights, labels=row.names(import),
           xlab="% of R-Square", pch=19,
           main="Relative Importance of Predictor Variables",
           sub=paste("Total R-Square=", round(rsquare, digits=3)),
           ...)
  return(import)
}
relbf(bf_step_back, col="blue")

##           Weights
## wrist    4.038431
## bicep    7.746760
## neck     8.238378
## hip     16.313788
## chest   21.795458
## abdomen 41.867184

The final model is:

\[ \hat{body-fat} = 1.52 -0.3965neck - 0.128chest + 1.01805abdomen -0.28758hip + 0.26bicep -1.55084wrist \]

and abdomen is relatively the most important predictor for predicting body fat percentage.

Looking at the $R^2$ value (multiple R-squared) from the summary output, 73.5% of the variability of density is explained by the regression on percentage of Height, Neck, Chest, Abdomen.

On average, holding the other variables constant, a 1 unit increase in Neck leads to a 0.3965 unit decrease in Body Fat Percentage
On average, holding the other variables constant, a 1 unit increase in Chest leads to a 0.128 unit decrease in Body Fat Percentage
On average, holding the other variables constant, a 1 unit increase in Abdomen leads to a 1.01805 unit increase in Body Fat Percentage
On average, holding the other variables constant, a 1 unit increase in Hip leads to a 0.28758 unit decrease in Body Fat Percentage
On average, holding the other variables constant, a 1 unit increase in Bicep leads to a 0.26 unit increase in Body Fat Percentage
On average, holding the other variables constant, a 1 unit increase in Wrist leads to a 1.5508 unit decrease in Body Fat Percentage

<<<<<<< HEAD

======= >>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419

2.2 BMI

For this analysis, the formula of BMI is

\[ BMI = \frac{Weight (lbs)*703}{Height(in)^2} \]

2.2.1 Defining the model with population parameters

\[ BMI = \beta_0 + \beta_1Neck + \beta_2Chest \\ + \beta_3Abdomen + \beta_4Waist + \beta_5Hip + \beta_6Thigh + \beta_7Knee + \beta_8Ankle \\+ \beta_9Bicep + \beta_{10}Forearm + \beta_{11}Wrist + \epsilon \]

qtlcharts::iplotCorr(data_bmi)

<<<<<<< HEAD ======= >>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419

Based on the interactive correlation matrix, it can be seen the level of correlation differs quite drastically between the variables and the backward variable selection method is also adopted here.

2.2.2 Multiple Regression with Variable Selection

bmi_lm = lm(bmi~.,data=data_bmi)
summary(bmi_lm)

## 
## Call:
## lm(formula = bmi ~ ., data = data_bmi)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -3.1790 -0.6548  0.0086  0.6881  3.8335 
## 
## Coefficients: (1 not defined because of singularities)
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept) -11.938661   1.725613  -6.919 4.19e-11 ***
## age           0.010049   0.007594   1.323   0.1870    
## neck          0.031474   0.056108   0.561   0.5754    
## chest         0.149956   0.022420   6.688 1.59e-10 ***
## abdomen       0.118223   0.020898   5.657 4.40e-08 ***
## waist               NA         NA      NA       NA    
## hip           0.060113   0.032231   1.865   0.0634 .  
## thigh         0.151883   0.034888   4.353 1.99e-05 ***
## knee         -0.265374   0.056116  -4.729 3.87e-06 ***
## ankle         0.067577   0.053692   1.259   0.2094    
## bicep         0.049712   0.041497   1.198   0.2321    
## forearm       0.086789   0.051015   1.701   0.0902 .  
## wrist        -0.045825   0.129082  -0.355   0.7229    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 1.058 on 238 degrees of freedom
## Multiple R-squared:  0.9046, Adjusted R-squared:  0.9002 
## F-statistic: 205.1 on 11 and 238 DF,  p-value: < 2.2e-16

Using the individual p-value method, the varaibles that need to be dropped are hip, ankle, bicep, forearm and wrist. To double check, the AIC criterion will also be considered.

bmi_step_back = step(bmi_lm, direction = "backward",trace = FALSE)
summary(bmi_step_back)

## 
## Call:
## lm(formula = bmi ~ chest + abdomen + hip + thigh + knee + forearm, 
##     data = data_bmi)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -3.1197 -0.6944 -0.0274  0.6831  3.8464 
## 
## Coefficients:
##              Estimate Std. Error t value Pr(>|t|)    
## (Intercept) -10.94257    1.43829  -7.608 6.10e-13 ***
## chest         0.16090    0.02122   7.582 7.18e-13 ***
## abdomen       0.12726    0.01826   6.968 3.01e-11 ***
## hip           0.05047    0.03084   1.637   0.1030    
## thigh         0.14983    0.03032   4.942 1.44e-06 ***
## knee         -0.23116    0.05148  -4.490 1.10e-05 ***
## forearm       0.11484    0.04468   2.571   0.0108 *  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 1.058 on 243 degrees of freedom
## Multiple R-squared:  0.9024, Adjusted R-squared:    0.9 
## F-statistic: 374.6 on 6 and 243 DF,  p-value: < 2.2e-16

Based on the backward selection model, the fitted model has become:

\[ \hat{BMI} = -10.94 +0.161Chest + 0.127Abdomen + 0.050Hip + 0.150 Thigh - 0.23Knee + 0.115Forearm \]

2.2.3 Check Assumptions

Finally, to check assumption, we perform the ggfortify function.

par(mfrow=c(1,2))
plot(bmi_step_back,which=1:2) + theme_bw()

<<<<<<< HEAD

=======

>>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419

## NULL

The QQ plot shows a straight line which indicates that the normality assumption is reasonable. However, the residuals vs fitted plot shows a fan shaped plot which indicates that the assumption of homogeneous variance is violated. We can use a log transformed response and re-fit the linear regression.

The new model will become: $log() = 1.83 +0.0058chest + 0.0052abdomen + 0.0064 thigh -0.0065knee + 0.0028bicep + 0.0040 forearm $.

ln_bmi_lm = lm(log(bmi)~.,data=data_bmi)
summary(ln_bmi_lm)

## 
## Call:
## lm(formula = log(bmi) ~ ., data = data_bmi)
## 
## Residuals:
##       Min        1Q    Median        3Q       Max 
## -0.133137 -0.024215 -0.000443  0.027919  0.102670 
## 
## Coefficients: (1 not defined because of singularities)
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  1.7512060  0.0654240  26.767  < 2e-16 ***
## age          0.0005158  0.0002879   1.791 0.074515 .  
## neck         0.0017502  0.0021272   0.823 0.411459    
## chest        0.0055041  0.0008500   6.475 5.37e-10 ***
## abdomen      0.0044728  0.0007923   5.645 4.68e-08 ***
## waist               NA         NA      NA       NA    
## hip          0.0010150  0.0012220   0.831 0.407026    
## thigh        0.0069490  0.0013227   5.253 3.31e-07 ***
## knee        -0.0082150  0.0021276  -3.861 0.000145 ***
## ankle        0.0026638  0.0020357   1.309 0.191940    
## bicep        0.0024437  0.0015733   1.553 0.121689    
## forearm      0.0038084  0.0019341   1.969 0.050112 .  
## wrist       -0.0017188  0.0048940  -0.351 0.725745    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.04009 on 238 degrees of freedom
## Multiple R-squared:  0.9079, Adjusted R-squared:  0.9036 
## F-statistic: 213.2 on 11 and 238 DF,  p-value: < 2.2e-16

ln_bmi_step_back = step(ln_bmi_lm, direction = "backward",trace = FALSE)
summary(ln_bmi_step_back)

## 
## Call:
## lm(formula = log(bmi) ~ age + chest + abdomen + thigh + knee + 
##     bicep + forearm, data = data_bmi)
## 
## Residuals:
##       Min        1Q    Median        3Q       Max 
## -0.132191 -0.023020 -0.000321  0.027716  0.106640 
## 
## Coefficients:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  1.7989069  0.0524420  34.303  < 2e-16 ***
## age          0.0004265  0.0002629   1.622 0.106083    
## chest        0.0058017  0.0008196   7.079 1.57e-11 ***
## abdomen      0.0047155  0.0007078   6.662 1.80e-10 ***
## thigh        0.0075055  0.0011966   6.273 1.62e-09 ***
## knee        -0.0069254  0.0018984  -3.648 0.000324 ***
## bicep        0.0026331  0.0015455   1.704 0.089720 .  
## forearm      0.0042579  0.0018372   2.318 0.021306 *  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.04002 on 242 degrees of freedom
## Multiple R-squared:  0.9067, Adjusted R-squared:  0.904 
## F-statistic: 335.8 on 7 and 242 DF,  p-value: < 2.2e-16

par(mfrow=c(1,2))
plot(ln_bmi_step_back,which=1:2) + theme_bw()

<<<<<<< HEAD

=======

>>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419

## NULL

sjPlot::tab_model(bmi_step_back, ln_bmi_step_back, digits = 5, show.ci = FALSE)

	bmi		log(bmi)
Predictors	Estimates	p	Estimates	p
(Intercept)	-10.94257	<0.001	1.79891	<0.001
chest	0.16090	<0.001	0.00580	<0.001
abdomen	0.12726	<0.001	0.00472	<0.001
hip	0.05047	0.103
thigh	0.14983	<0.001	0.00751	<0.001
knee	-0.23116	<0.001	-0.00693	<0.001
forearm	0.11484	0.011	0.00426	0.021
age			0.00043	0.106
bicep			0.00263	0.090
Observations	250		250
R² / R² adjusted	0.902 / 0.900		0.907 / 0.904

However, although the transformation has aided with the homogeneous variance assumption, the interpretation itself does not make much sense - BMI is determined by an increase in value, not the increase in percentage change. Hence in the final comparison, we will use the untransformed model.

2.2.4 Final Fitted Model

relbmi <- function(fit,...){
  R <- cor(fit$model)
  nvar <- ncol(R)
  rxx <- R[2:nvar, 2:nvar]
  rxy <- R[2:nvar, 1]
  svd <- eigen(rxx)
  evec <- svd$vectors
  ev <- svd$values
  delta <- diag(sqrt(ev))
  lambda <- evec %*% delta %*% t(evec)
  lambdasq <- lambda ^ 2
  beta <- solve(lambda) %*% rxy
  rsquare <- colSums(beta ^ 2)
  rawwgt <- lambdasq %*% beta ^ 2
  import <- (rawwgt / rsquare) * 100
  import <- as.data.frame(import)
  row.names(import) <- names(fit$model[2:nvar])
  names(import) <- "Weights"
  import <- import[order(import),1, drop=FALSE]
  dotchart(import$Weights, labels=row.names(import),
           xlab="% of R-Square", pch=19,
           main="Relative Importance of Predictor Variables",
           sub=paste("Total R-Square=", round(rsquare, digits=3)),
           ...)
  return(import)
}
relbmi(bmi_step_back, col="blue")

##           Weights
## forearm  9.021850
## knee     9.080285
## thigh   15.075099
## hip     17.245099
## abdomen 24.705834
## chest   24.871833

The final model is:

\[\hat{BMI} = -10.94 +0.161Chest + 0.127Abdomen + 0.050Hip + 0.150 Thigh - 0.23Knee + 0.115Forearm \] and both chest and abdomen are relatively more important in predicting BMI.

Looking at the $R^2$ value (multiple R-squared) from the summary output, 90.2% of the variability of density is explained by the regression on percentage of Height, Neck, Chest, Abdomen.

On average, holding the other variables constant, a 1 unit increase in Chest leads to a 0.161 unit increase in BMI
On average, holding the other variables constant, a 1 unit increase in Abdomen leads to a 0.127 unit increase in BMI
On average, holding the other variables constant, a 1 unit increase in Hip leads to a 0.050 unit increase in BMI
On average, holding the other variables constant, a 1 unit increase in Thigh leads to a 0.15 unit increase in BMI
On average, holding the other variables constant, a 1 unit increase in Knee leads to a 0.23 unit decrease in BMI
On average, holding the other variables constant, a 1 unit increase in Forearm leads to a 0.115 unit decrease in BMI

2.3 Body Density

2.3.1 Defining the model with population parameters

\[ Body Density = \beta_0 + \beta_1Pcf.BF + \beta_2Age + \beta_3Weight + \beta_4Height\\ + \beta_5Neck + \beta_6Chest + \beta_7Abdomen + \beta_8Waist + \beta_9Hip + \beta_{10}Thigh\\ + \beta_{11}Knee + \beta_{12}Ankle + \beta_{13}Bicep + \beta_{14}Forearm + \beta_{15}Wrist + \epsilon \]

#data1<-data_density[,-2]

cor_matrix <- cor(data_density)
pheatmap(cor_matrix, display_numbers = T,na.rm=T)

<<<<<<< HEAD

=======

>>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419

Above matrix has shown the interactice correlation between variables. Notbaly, Pct.BF has a -0.99 relationship with Density, which means Pct.BF could be used to explain Density. Meanwhile, variables having similar properties are linked together, which could be useful for generating groups.

2.3.2 Check Assumptions

The residuals $\epsilon_i$ are iid $N(0,\sigma^2)$ and there is a linear relationship between y and x.

M0 <- lm(density ~ 1, data = data_density)  # Null model
M1 <- lm(density ~ ., data = data_density)  # Full model
autoplot(M1,which=1:2)+theme_bw()

<<<<<<< HEAD

=======

>>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419

round(summary(M1)$coef, 3)

##             Estimate Std. Error t value Pr(>|t|)
## (Intercept)    1.082      0.016  68.068    0.000
## neck           0.001      0.001   2.074    0.039
## chest          0.000      0.000   1.818    0.070
## abdomen       -0.002      0.000 -13.645    0.000
## hip            0.001      0.000   2.950    0.003
## thigh          0.000      0.000  -0.980    0.328
## knee           0.000      0.001   0.689    0.492
## ankle          0.000      0.001  -0.675    0.500
## bicep         -0.001      0.000  -1.357    0.176
## forearm        0.000      0.000  -0.947    0.345
## wrist          0.004      0.001   3.310    0.001

step.fwd.aic <- step(M0, scope = list(lower = M0, upper = M1), direction = "forward", trace = FALSE)
summary(step.fwd.aic)

## 
## Call:
## lm(formula = density ~ waist + wrist + hip + chest + bicep + 
##     neck, data = data_density)
## 
## Residuals:
##       Min        1Q    Median        3Q       Max 
## -0.024142 -0.007680  0.000523  0.006156  0.038390 
## 
## Coefficients:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  1.0844660  0.0154768  70.070  < 2e-16 ***
## waist       -0.0060402  0.0004401 -13.724  < 2e-16 ***
## wrist        0.0038812  0.0010612   3.657 0.000312 ***
## hip          0.0006990  0.0002153   3.247 0.001331 ** 
## chest        0.0003881  0.0002097   1.851 0.065427 .  
## bicep       -0.0007779  0.0003535  -2.201 0.028695 *  
## neck         0.0009609  0.0005185   1.853 0.065030 .  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.01007 on 243 degrees of freedom
## Multiple R-squared:  0.722,  Adjusted R-squared:  0.7152 
## F-statistic: 105.2 on 6 and 243 DF,  p-value: < 2.2e-16

step.back.aic <- step(M1, scope = list(lower = M0, upper = M1), direction = "backward", trace = FALSE)
summary(step.back.aic)

## 
## Call:
## lm(formula = density ~ neck + chest + abdomen + hip + bicep + 
##     wrist, data = data_density)
## 
## Residuals:
##       Min        1Q    Median        3Q       Max 
## -0.024142 -0.007680  0.000523  0.006156  0.038390 
## 
## Coefficients:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  1.0844660  0.0154768  70.070  < 2e-16 ***
## neck         0.0009609  0.0005185   1.853 0.065030 .  
## chest        0.0003881  0.0002097   1.851 0.065428 .  
## abdomen     -0.0023780  0.0001733 -13.724  < 2e-16 ***
## hip          0.0006990  0.0002153   3.247 0.001331 ** 
## bicep       -0.0007779  0.0003535  -2.201 0.028694 *  
## wrist        0.0038812  0.0010612   3.657 0.000312 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.01007 on 243 degrees of freedom
## Multiple R-squared:  0.722,  Adjusted R-squared:  0.7152 
## F-statistic: 105.2 on 6 and 243 DF,  p-value: < 2.2e-16

exh <- regsubsets(density~., data = data_density, nvmax = 15)

## Warning in leaps.exhaustive(a, really.big): XHAUST returned error code -999

plot(exh,scale="bic")

<<<<<<< HEAD

=======

>>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419

2.3.3 Multiple Regression using the BIC

M2<- lm(formula = density ~ neck + chest + abdomen, 
    data = data_density)
summary(M2)

## 
## Call:
## lm(formula = density ~ neck + chest + abdomen, data = data_density)
## 
## Residuals:
##       Min        1Q    Median        3Q       Max 
## -0.029395 -0.007156 -0.000682  0.007305  0.046687 
## 
## Coefficients:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  1.1392164  0.0119499  95.333  < 2e-16 ***
## neck         0.0017212  0.0004599   3.743 0.000226 ***
## chest        0.0004569  0.0002135   2.140 0.033361 *  
## abdomen     -0.0021095  0.0001592 -13.250  < 2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.01057 on 246 degrees of freedom
## Multiple R-squared:  0.6904, Adjusted R-squared:  0.6867 
## F-statistic: 182.9 on 3 and 246 DF,  p-value: < 2.2e-16

M3<- lm(formula = density ~ neck + chest + abdomen + waist , 
    data = data_density)
summary(M3)

## 
## Call:
## lm(formula = density ~ neck + chest + abdomen + waist, data = data_density)
## 
## Residuals:
##       Min        1Q    Median        3Q       Max 
## -0.029395 -0.007156 -0.000682  0.007305  0.046687 
## 
## Coefficients: (1 not defined because of singularities)
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  1.1392164  0.0119499  95.333  < 2e-16 ***
## neck         0.0017212  0.0004599   3.743 0.000226 ***
## chest        0.0004569  0.0002135   2.140 0.033361 *  
## abdomen     -0.0021095  0.0001592 -13.250  < 2e-16 ***
## waist               NA         NA      NA       NA    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.01057 on 246 degrees of freedom
## Multiple R-squared:  0.6904, Adjusted R-squared:  0.6867 
## F-statistic: 182.9 on 3 and 246 DF,  p-value: < 2.2e-16

Drop waist and add other variables

M4<- lm(formula = density ~ neck + chest + abdomen + hip , 
    data = data_density)
summary(M4)

## 
## Call:
## lm(formula = density ~ neck + chest + abdomen + hip, data = data_density)
## 
## Residuals:
##       Min        1Q    Median        3Q       Max 
## -0.028989 -0.007256  0.000047  0.006767  0.045116 
## 
## Coefficients:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  1.1150223  0.0139695  79.818  < 2e-16 ***
## neck         0.0014682  0.0004584   3.203  0.00154 ** 
## chest        0.0003734  0.0002113   1.768  0.07837 .  
## abdomen     -0.0023671  0.0001759 -13.455  < 2e-16 ***
## hip          0.0006619  0.0002074   3.191  0.00160 ** 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.01037 on 245 degrees of freedom
## Multiple R-squared:  0.7028, Adjusted R-squared:  0.6979 
## F-statistic: 144.8 on 4 and 245 DF,  p-value: < 2.2e-16

M5<- lm(formula = density ~ neck + chest + abdomen + hip + thigh , 
    data = data_density)
summary(M5)

## 
## Call:
## lm(formula = density ~ neck + chest + abdomen + hip + thigh, 
##     data = data_density)
## 
## Residuals:
##       Min        1Q    Median        3Q       Max 
## -0.029545 -0.006988  0.000516  0.007098  0.043367 
## 
## Coefficients:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  1.1073577  0.0144127  76.832  < 2e-16 ***
## neck         0.0016460  0.0004644   3.544 0.000472 ***
## chest        0.0003393  0.0002107   1.610 0.108626    
## abdomen     -0.0023869  0.0001752 -13.627  < 2e-16 ***
## hip          0.0010639  0.0002889   3.682 0.000285 ***
## thigh       -0.0005716  0.0002878  -1.986 0.048171 *  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.01031 on 244 degrees of freedom
## Multiple R-squared:  0.7075, Adjusted R-squared:  0.7015 
## F-statistic:   118 on 5 and 244 DF,  p-value: < 2.2e-16

Drop chest and add other variables

M6<- lm(formula = density ~ neck + abdomen + hip + thigh + knee , 
    data = data_density)
summary(M6)

## 
## Call:
## lm(formula = density ~ neck + abdomen + hip + thigh + knee, data = data_density)
## 
## Residuals:
##       Min        1Q    Median        3Q       Max 
## -0.029361 -0.007760  0.000360  0.007152  0.043816 
## 
## Coefficients:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  1.1035095  0.0150104  73.517  < 2e-16 ***
## neck         0.0018149  0.0004395   4.129    5e-05 ***
## abdomen     -0.0022098  0.0001347 -16.402  < 2e-16 ***
## hip          0.0010035  0.0002984   3.363 0.000894 ***
## thigh       -0.0007038  0.0002938  -2.395 0.017355 *  
## knee         0.0007551  0.0005006   1.508 0.132778    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.01032 on 244 degrees of freedom
## Multiple R-squared:  0.7071, Adjusted R-squared:  0.7011 
## F-statistic: 117.8 on 5 and 244 DF,  p-value: < 2.2e-16

Drop knee

M7<- lm(formula = density ~ neck + abdomen + hip + thigh , 
    data = data_density)
summary(M7)

## 
## Call:
## lm(formula = density ~ neck + abdomen + hip + thigh, data = data_density)
## 
## Residuals:
##       Min        1Q    Median        3Q       Max 
## -0.030204 -0.007301  0.000653  0.007199  0.045107 
## 
## Coefficients:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  1.1104052  0.0143343  77.465  < 2e-16 ***
## neck         0.0019085  0.0004362   4.375  1.8e-05 ***
## abdomen     -0.0022064  0.0001351 -16.337  < 2e-16 ***
## hip          0.0011314  0.0002868   3.945 0.000104 ***
## thigh       -0.0006094  0.0002878  -2.117 0.035236 *  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 0.01035 on 245 degrees of freedom
## Multiple R-squared:  0.7044, Adjusted R-squared:  0.6996 
## F-statistic:   146 on 4 and 245 DF,  p-value: < 2.2e-16

2.3.4 Final Fitted Model

relbmi <- function(fit,...){
  R <- cor(fit$model)
  nvar <- ncol(R)
  rxx <- R[2:nvar, 2:nvar]
  rxy <- R[2:nvar, 1]
  svd <- eigen(rxx)
  evec <- svd$vectors
  ev <- svd$values
  delta <- diag(sqrt(ev))
  lambda <- evec %*% delta %*% t(evec)
  lambdasq <- lambda ^ 2
  beta <- solve(lambda) %*% rxy
  rsquare <- colSums(beta ^ 2)
  rawwgt <- lambdasq %*% beta ^ 2
  import <- (rawwgt / rsquare) * 100
  import <- as.data.frame(import)
  row.names(import) <- names(fit$model[2:nvar])
  names(import) <- "Weights"
  import <- import[order(import),1, drop=FALSE]
  dotchart(import$Weights, labels=row.names(import),
           xlab="% of R-Square", pch=19,
           main="Relative Importance of Predictor Variables",
           sub=paste("Total R-Square=", round(rsquare, digits=3)),
           ...)
  return(import)
}
<<<<<<< HEAD
relbmi(bmi_step_back, col="blue")

======= relweights(bmi_step_back, col="blue")

>>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419

##           Weights
## forearm  9.021850
## knee     9.080285
## thigh   15.075099
## hip     17.245099
## abdomen 24.705834
## chest   24.871833

Obviously, abdomen contributes the most in the relationship with body density.

The final model is: \[ Body Density = 1.1104052 + 0.0019085 \times Neck\\ - 0.0022064 \times Abdomen\ + 0.0011314 \times Hip\\ - 0.0006094 \times Thigh\\ \] Looking at the $R^2$ value (multiple R-squared) from the summary output, 70.4% of the variability of density is explained by the regression on percentage of Height, Neck, Chest, Abdomen.

On average, holding the other variables constant, a 1% increase in Neck leads to a 0.2% unit increase in Density
On average, holding the other variables constant, a 1% increase in Abdomen leads to a 0.2% decrease in Density
On average, holding the other variables constant, a 1% increase in Hip leads to a 0.1% increase in Density
On average, holding the other variables constant, a 1% increase in Thigh leads to a 0.06% decrease in Density

2.3.5 Linear regression assumptions for the stepwise model

autoplot(M7,which=1:2)+theme_bw()

<<<<<<< HEAD

=======

>>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419

2.4 Which variables can best predict whether a person is overweight?

Since the variable ‘Overweight’ is dichotomous (binary), we perform a logistical regression to interpret the relationship between overweight and other significant variables.

2.4.1 Checking for Significance in a Logistic Regression

data_overweight = data[-c(4:5,17)]
data_overweight$overweight = as.numeric(data_overweight$overweight)
glm1 = glm(overweight ~ ., data = data_overweight)
# drop knee
glm2 = glm(overweight ~ density + pct_bf + age + neck + chest + abdomen + waist + hip + thigh + ankle + bicep + forearm + wrist, data = data_overweight)
# drop ankle
glm3 = glm(overweight ~ density + pct_bf + age + neck + chest + abdomen + waist + hip + thigh + bicep + forearm + wrist, data = data_overweight)
# drop density
glm4 = glm(overweight ~ pct_bf + age + neck + chest + abdomen + waist + hip + thigh + bicep + forearm + wrist, data = data_overweight)
# drop age
glm5 = glm(overweight ~ pct_bf + neck + chest + abdomen + waist + hip + thigh + bicep + forearm + wrist, data = data_overweight)
# drop waist
glm6 = glm(overweight ~ pct_bf + neck + chest + abdomen + hip + thigh + bicep + forearm + wrist, data = data_overweight)
# drop neck
glm7 = glm(overweight ~ pct_bf + chest + abdomen + hip + thigh + bicep + forearm + wrist, data = data_overweight)
# drop pct_bf
glm8 = glm(overweight ~ chest + abdomen + hip + thigh + bicep + forearm + wrist, data = data_overweight)
# drop waist
glm9 = glm(overweight ~ chest + abdomen + hip + thigh + bicep + forearm, data = data_overweight)
# drop wrist
glm10 = glm(overweight ~ chest + abdomen + hip + thigh + bicep + forearm, data = data_overweight)
# drop forearm
glm11 = glm(overweight ~ chest + abdomen + hip + thigh + bicep, data = data_overweight)
# drop thigh
glm12 = glm(overweight ~ chest + abdomen + hip + bicep, data = data_overweight)
# drop hip
glm13 = glm(overweight ~ chest + abdomen + bicep, data = data_overweight)
summary(glm13)

## 
## Call:
## glm(formula = overweight ~ chest + abdomen + bicep, data = data_overweight)
## 
## Deviance Residuals: 
##      Min        1Q    Median        3Q       Max  
## -0.69987  -0.25440  -0.02293   0.26254   0.68701  
## 
## Coefficients:
##              Estimate Std. Error t value Pr(>|t|)    
## (Intercept) -3.845275   0.282551 -13.609  < 2e-16 ***
## chest        0.013281   0.006416   2.070 0.039503 *  
## abdomen      0.020164   0.004805   4.196 3.79e-05 ***
## bicep        0.035618   0.009834   3.622 0.000355 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## (Dispersion parameter for gaussian family taken to be 0.1026899)
## 
##     Null deviance: 62.500  on 249  degrees of freedom
## Residual deviance: 25.262  on 246  degrees of freedom
## AIC: 146.43
## 
## Number of Fisher Scoring iterations: 2

Before we start making predictions with the model, we drop the variables which are not a significant predictor for being overweight. The fitted model is shown below.

2.4.2 Fitted Model

\[ logit(p) = log(\frac{p}{1-p}) = -3.845275 + 0.013281 \times Chest\\ + 0.020164 \times Abdomen\ + 0.035618 \times bicep\\ \] where the logit(p) is a special link from our linear combination of predictors to the probability of the outcome being equal to 1, and the coefficients are interpreted as changes in log-odds.

2.4.3 Visualisation and Output of the model coefficients (odds scale)

tab_model(glm13)

	overweight
Predictors	Estimates	CI	p
(Intercept)	-3.85	-4.40 – -3.29	<0.001
chest	0.01	0.00 – 0.03	0.040
abdomen	0.02	0.01 – 0.03	<0.001
bicep	0.04	0.02 – 0.05	<0.001
Observations	250
R² Nagelkerke	0.626

plot_model(glm13) + theme_bw(base_size = 10) + ylim(0, 0.10) + labs(x = "Overweight", y = "Odds Ratios", title = "Model Coefficients using odds scale")

plot_model(glm13, type = "pred", terms = c("abdomen", "chest", "bicep"), show.data = TRUE) +
  theme_bw(base_size = 10)

In the Coefficient graph, the three significant variables have similar odd ratios giving the model a smaller confidence interval.

From the Prediction graph, the positive slopes in the highlighted areas indicate a larger abdomen and chest circumference leads to a high probability of being overweight. Comparing the three individual graphs, the slight difference in the slope’s y axis indicates bicep circumference correlates with odds of being overweight.

2.4.4 Predictions

We correctly classified 91.2% of the observations, hence our resubstitution error rate, proportion of data predicted incorrectly using the fitted model, is 8.8%.

data$overweight = as.numeric(data$overweight)
glm0 = glm(overweight ~ chest + abdomen + bicep, data = data)
data = data %>% 
  mutate(pred_prob = predict(glm0, type = "response"),
         pred_surv = round(pred_prob))
mean(data$overweight == data$pred_surv)

## [1] 0.912

The odds of being overweight for someone with an above average abdomen circumference of 120 is 1.26.

predict_overweight = data.frame(abdomen = 130, chest = mean(data$chest), bicep = mean(data$bicep))
predict(glm13, newdata = predict_overweight, type = "link")

##        1 
## 1.260437

The odds of being overweight for someone with a below average abdomen circumference of 60 is -0.15.

predict_overweight = data.frame(abdomen = 60, chest = mean(data$chest), bicep = mean(data$bicep))
predict(glm13, newdata = predict_overweight, type = "link")

##          1 
## -0.1510207

The odds of being overweight for someone with above average circumferences for their abdomen, chest, and bicep is 1.37.

predict_overweight = data.frame(abdomen = mean(data$abdomen)*1.2, chest = mean(data$chest)*1.2, bicep = mean(data$bicep)*1.2)
predict(glm13, newdata = predict_overweight, type = "link")

##        1 
## 1.369055

<<<<<<< HEAD

DECISION TREE NOT WORKING delete??

library(partykit)

## Warning: package 'partykit' was built under R version 3.5.2

## Warning: package 'libcoin' was built under R version 3.5.2

## Warning: package 'mvtnorm' was built under R version 3.5.2

library(rpart)
data_overweight$overweight=as.character(data_overweight$overweight)
ov_tree = rpart(overweight ~ abdomen + chest + bicep, data = data_overweight, method = "class",control = rpart.control(cp = 0.009))
=======

Decision Tree
data_overweight$overweight=as.character(data_overweight$overweight)

#library(caret)
#confusion.glm = confusionMatrix(
#  data = as.factor(data$pred_surv), 
#  reference = as.factor(data$overweight))
#confusion.glm$table
ov_tree = rpart(overweight ~ abdomen + chest + bicep, data = data_overweight, method = "class",control = rpart.control(cp = 0.009))
>>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419
ov_tree
## n= 250 
## 
## node), split, n, loss, yval, (yprob)
##       * denotes terminal node
## 
## 1) root 250 125 0 (0.50000000 0.50000000)  
##   2) chest< 101.55 144  23 0 (0.84027778 0.15972222)  
##     4) abdomen< 92.6 126  10 0 (0.92063492 0.07936508) *
##     5) abdomen>=92.6 18   5 1 (0.27777778 0.72222222) *
##   3) chest>=101.55 106   4 1 (0.03773585 0.96226415) *
plot(as.party(ov_tree))
<<<<<<< HEAD

=======

>>>>>>> 3b77f7033bba3b3a34a01a68e0cb0ab306ff7419



2.5.1 Interpretation



3. Limitations


Gender Bias


Privacy Issues


Age Range






4. Conclusion


5. References


Australian Institute of Health and Welfare (AIHW). (2019). Overweight & obesity. Australian Government. Retrieved from https://www.aihw.gov.au/reports-data/behaviours-risk-factors/overweight-obesity/overview


DASL. (n.d.). Bodyfat. DASL. Retrieved from < https://dasl.datadescription.com/datafile/bodyfat>


Rahman, A., & Harding, A. (2013). Prevalence of overweight and obesity epidemic in Australia: some causes and consequences. JP Journal of Biostatistics, 10(1), 31-48.


Stein, C. J., & Colditz, G. A. (2004). The epidemic of obesity. The Journal of Clinical Endocrinology & Metabolism, 89(6), 2522-2525.

DATA2002 Module 4 Report

Group 9 (460352996,480407614,480145820,470066919)

1. Introduction

1.1 Sampling method and potential biases

1.2 Data import, processing and cleaning

1.3 Data Visualisation

1.3.1 Obesity and Age

1.3.2 Body Measurements

1.4 Analysis Approach

2. Analysis

2.1 Body Fat Percentage

2.1.1 Data Visualisation

2.1 Body Fat Percentage

2.1.1 Data Visualisation

2.1.2 Multiple Regression and Variable Selection

2.1.3 Check Assumptions

2.1.4 Final fitted model

2.2 BMI

2.2.1 Defining the model with population parameters

2.2.2 Multiple Regression with Variable Selection

2.2.3 Check Assumptions

2.2.4 Final Fitted Model

2.3 Body Density

2.3.1 Defining the model with population parameters

2.3.2 Check Assumptions

2.3.3 Multiple Regression using the BIC

2.3.4 Final Fitted Model

2.3.5 Linear regression assumptions for the stepwise model

2.4 Which variables can best predict whether a person is overweight?

2.4.1 Checking for Significance in a Logistic Regression

2.4.2 Fitted Model

2.4.3 Visualisation and Output of the model coefficients (odds scale)

2.4.4 Predictions

DECISION TREE NOT WORKING delete??

Decision Tree

2.5.1 Interpretation

3. Limitations

4. Conclusion

5. References